School of Electrical and Electronic Engineering,

1. Real-Time Power Quality DisturbancesClassification Using Hybrid Method Based onS-Transform and Dynamics Li Kaicheng,He Shunfan, Ming Zhang School of Electrical and Electronic Engineering, Huazhong University of Science and Technology, Wuhan, China E-mail：likaicheng@hust.edu.cn

2. Power Quality Problems New energies are widely used Diverse nonlinear loads charging pile • intermediate frequency furnace electric arc furnace

3. Power Quality Events

4. How to classify PQDs How to realize Power Quality DisturbancesClassification with real time? Accurate: better than other methods Quick: meet the real time requirements

5. ABSTRACT This paper proposes a real-time power quality disturbances (PQDs) classification by using a hybrid method (HM) based on S-transform (ST) and dynamics (Dyn). Classification accuracy and run time are distinctly improved. The HM uses Dyn to identify the location of the signal components in the frequency spectrum yielded by Fourier transform, and uses inverse Fourier transform to only some of the signal components. Then features from Fourier transform, ST,and Dyn are selected, A decision tree is used to classify the types of PQD. In order to reduce the influence of Heisenberg’s uncertainty, we proposed that different signal components are windowed by different Gaussian windows, which brings better adaption and flexibility. Simulation and test show that the run time of the application has been greatly reduced with satisfactory classification accuracy.

6. Introduction 1.Feature extraction 2.Categorization PQDs classification steps

7. Traditional Method-Feature extraction • Wavelet Transform-WT or Wavelet Pocket Transform-WPT Wavelet features extraction relies on the number of decomposed level and wavelet basis. The frequency amplitude could not be counted.This disadvantage always perplexes the classifier。 • S-transform (ST) ST has a high tolerance of noise and guarantees satisfactory accuracy on classification.The main disadvantage of ST isheavy computation burden. • Kalman filter (KF) KF-based approaches depend wholly on the model of state and measurement matrixes. Any mismatch of signal and filter model will bring errors.

8. Traditional Method-Classifier • Artificial neural network (ANN) ANN can be trained online and refreshed by each new sample. • support vector machine (SVM) SVM has a strong learning process and proved useful when the sample group is small. • fuzzy logic (FL) and decision tree (DT) FL and DT do not need training and rely on rules. • ProblemMost of reported literatures can hardly be employed for real-timeapplication and even many of them are only discussed insimulations.

9. The hybrid method Dynamics+ST It makes use of Dyn to reduce the run time of ST significantly and extracts five distinctive features with satisfactory classification accuracy. In the consideration of different uses, DT is adopted as theclassifier for better compatibility. This paper investigates 8 single PQDs and 2 complexPQDs including: harmonic, flicker, sag, interruption, swell,spike, multiple notch, oscillation, sag plus harmonic, and swell plus harmonic

10. Dynamics (Dyn) Dynamics (Dyn) was firstly introduced by Grimaud in1992. It is employed to find the extreme points in a sequence under noises. The algorithm is easy and with low computation burden. • Brief review of dynamics: • Define a path P(m, n), where m and n are the endpoints of P, Dyn of the path is the difference of top point and button point of the path. Then the Dyn is where Sup is the supremum, pi and pj are the highest point and lowest point, respectively, in path P, and halt(.) is the altitude of the point.

11. Dynamics (Dyn) Example: noise elimination by using Dyn. A path y = f (x) is shown in Fig. 1. M1 and P1 are two of maximal points: Dynmax(M1) = Dyn(M1, N1) Dynmax(P1) = Dyn(O1, P1). If Dyn(O1, P1) < Inf <Dyn(M1, N1), point M1 will be marked by Dyn and point P1 will be labeled with noise and eliminated. where Inf is the infimum.

12. Dynamics (Dyn) In Fig. 1, M2 and P2 are two minimal points: Dynmin(M2) = Dyn(L2, M2) Dynmin(P2) = Dyn(P2, Q2). If Dyn(P2, Q2) < Inf <Dyn(L2, M2), point P2 will be deemed as noise. Then all the points indicated by Dyn are marked with red stars.

13. S-Transform ------Gaussian window Time-frequency analysis method; Good resolution . DFT Discrete form of S transform 2) Brief review of S-transform:

14. S-Transform • The process of ST can be generally divided into four steps.Step 1: Use fast Fourier transform (FFT) to the input signaland obtain the spectrum H(m) where m is the frequency sample index.Step 2: Shift H(m) with n.Step 3: Compute the FFT of Gaussian window where a is the parameter to tune the shape of Gaussian window. • Step 4: Multiply each H(m + n) with the correspondingG(m, n) and use IFFT to the result. Then the discreteST is obtained as • where, N point number, T is sampling interval

15. The hybrid method In ST, it should be noted that each H(m+n) has its own Gaussian window. The main computation burden of ST is that every H(m + n) is used in steps 3 and 4. Most of frequency samples yielded by FFT are 0 or just noises which have few information of the signal. Hence, the key to reduce run time is to find the location of frequency samples which are not 0 and noises in the spectrum. By using Dyn, the interested frequency samples are selected from 0 and noises in the spectrum and only some of them are used in steps 3 and 4.

16. The hybrid method Gaussian window plays an important role in ST. In low frequency band, high time resolution is expected for fast response of waveform changes. In high frequency band, high frequency resolution is expected for good tolerance of noises and influence from other frequency components. According to Heisenberg’s uncertainty, time and frequency resolution can not increase at the same time. From (4), better time resolution comes with a < 1, while better frequency resolution comes with a > 1. In this paper, we used G1(a < 1) and G2(a > 1) to address low frequency components (f ≤ 350 Hz) and high frequency ones (f > 350 Hz), respectively. By many experiments, G1(a = 0.8) and G2(a = 10) give good performances.

17. The hybrid method

18. The hybrid method

19. Features Extraction

20. B. Features Extraction • C1: The mean value of fundamental amplitude. Parameter d is the threshold of noise tolerance. C1 is used to differentiate spike and multiple notch from noise. • C2: The sum number of Dynmax and Dynmin of fundamental curve. It can be used to recognize flicker. • C3: Total harmonic distortion (THD). The THD is obtained by FFT. Each harmonic peak is indicated by If THD > 5%, C3 = 1 else, C3 = 0. • C4: C4 is used to identify oscillation.The number of Dynmax of high frequency component curves. If there is one Dyn in one of the curves, C4 is 1, else C4 is 0. • C5: The amplitude of the point indicated by C2, when C2 is 1. If 0.1 < C5 < 0.9, then it is sag; if 0 < C5 < 0.1, then it is interrupt; if 1.1 < C5 < 1.8, then it is swell. .

21. Rules for Decision Tree we proposed a common use of PQDs classification. The rules of DT are presented in Table I. Every node of the tree needs only one feature rather than a group of features.

22. Rules for Decision Tree 4 times (normal, harmonic) 3 times(sag, swell, sag plus harmonic, swell plus harmonic, oscillation), 2 times(spike, multiple notch, interrupt), 1 time(flicker). The average comparing times is:

23. Simulations The formulas of simulated signals are mainly referred from: C. Lee and Y. Shen, “Optimal feature selection for power quality disturbances classification,” IEEE Trans. Power Del., vol. 26, no. 4,pp. 1250–1257, Oct. 2011. Signal contaminated with 30 dB noise. Dynmax is used in Fourier spectrum, fundamental curve and high frequency component curves. Dynmin is used in fundamental curve. The noise threshold d is 0.005 by many experiments. Run time of ST mainly depends on the number of IFFT. Thus,the efficiency of HM can be evaluated by this number. The sample rate is 5 kHz and the length of all simulated signals isten cycles (0.2 s), which means one time ST needs 500 IFFTs.

24. Simulations Figs. (a) are the signals. Figs. (b) are the fundamental curves. Figs. (c) are the spectrum of the signals yielded by FFT. Figs. (d) are the high frequency component curves. All points indicated by Dyn are marked with red stars. Run time of HM in this group of simulations compared to the one of ST is shown in Table II by assuming the run time of ST is 1 s. It can be seen that the run time has been reduced greatly. In a certain length of signal, higher sample rate brings higher efficiency of HM.

25. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Flicker signal and its features

26. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Oscillation signal and its features

27. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Multiple notch signal and its features

28. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Sag + harmonic signal and its features

29. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Spike signal and its features

30. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Swell + harmonic signal and its features

31. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Sag signal and its features

32. Simulations Fig. (a) is the signal. Fig. (b) is the fundamental curve. Fig. (c) is the spectrum of the signal yielded by FFT. Fig. (d) is the high frequency component curve. Swell signal and its features

33. Simulations--HM versus M1and M2 • we generated randomly 100 signals of each type of PQDs which are contaminated with 40, 30, and 20 dB noise, respectively. Since HM is based on ST and has real time ability, two related methods are adopted to make comparisons: method in M1 (M1) F. Zhao and R. Yang, “Power quality disturbance recognition using Stransform,” IEEE Trans. Power Del., vol. 22, no. 2, pp. 944–207, Apr.2007 • and method in M2. (M2) M. Zhang, K. Li, and Y. Hu, “A real time classification method forpower quality disturbances,” Electr. Power Syst. Res., vol. 81, no. 1, pp.660–666, 2011 The classification results of HM, M1, and M2 are presented in Table III.

34. HM versus M2

35. Experiment A hardware platform is introduced to runHM. Laboratory experiments are designed to evaluate the run time and test the correctness of HM by using real standard signals. Field signal analyses are also presented to evaluate the performance of HM in industry application,

36. System Structure • Hardware system structure

37. Laboratory Experiments Fluke 6100A is used as a standard signal source The personal computer is adopted to debug the routineand display the results in Code Composer Studio. The length of tested signals is ten cycles (0.2 s). From Fig. 3 andTable II, flicker takes the shortest run time while harmonictakes the longest one among the PQDs.

38. Laboratory Experiments Harmonic:run time 0.072 s Flicker:run time 0.038 s (a) Test performance of flicker on DSP. (b) Test performance of harmonic on DSP

39. Thank you !